[Mesa-dev] [PATCH v5 08/34] nvir/nir: run some passes to make the conversion easier

[Pierre Moreau]

On 2018-02-20 — 22:02, Karol Herbst wrote:
> v2: add constant_folding
> 
> Signed-off-by: Karol Herbst <kherbst at redhat.com>
> ---
>  .../drivers/nouveau/codegen/nv50_ir_from_nir.cpp   | 40 ++++++++++++++++++++++
>  1 file changed, 40 insertions(+)
> 
> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp
> index 73527d4800..148db464bd 100644
> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp
> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_nir.cpp
> @@ -29,6 +29,12 @@
>  #include "codegen/nv50_ir_lowering_helper.h"
>  #include "codegen/nv50_ir_util.h"
>  
> +static int
> +type_size(const struct glsl_type *type)

From looking at the implementation of glsl_count_attribute_slots, it seems to
return the number of elements more than the number of bits/bytes taken up by
the type. So I should change the name of this function.

> +{
> +   return glsl_count_attribute_slots(type, false);
> +}
> +
>  namespace {
>  
>  using namespace nv50_ir;
> @@ -50,6 +56,40 @@ Converter::Converter(Program *prog, nir_shader *nir, nv50_ir_prog_info *info)
>  bool
>  Converter::run()
>  {
> +   bool progress;
> +
> +   if (prog->dbgFlags & NV50_IR_DEBUG_BASIC)
> +      nir_print_shader(nir, stderr);
> +
> +   // converts intrinsic load_var to intrinsic load_uniform
> +   NIR_PASS_V(nir, nir_lower_io, nir_var_all, type_size, (nir_lower_io_options)0);
> +
> +   NIR_PASS_V(nir, nir_lower_regs_to_ssa);
> +   NIR_PASS_V(nir, nir_lower_load_const_to_scalar);
> +
> +   do {
> +      progress = false;
> +      // we need this to_ssa otherwise the later opts are less effective
> +      NIR_PASS_V(nir, nir_lower_vars_to_ssa);

Do you need to lower to SSA on every loop run, shouldn’t it be enough to only
run it once?

> +      NIR_PASS(progress, nir, nir_lower_alu_to_scalar);
> +      NIR_PASS(progress, nir, nir_lower_phis_to_scalar);
> +      // some ops depend on having constant as sources, but those can also
> +      // point to expressions made from constants like 0 + 1
> +      NIR_PASS(progress, nir, nir_opt_constant_folding);
> +      NIR_PASS(progress, nir, nir_copy_prop);
> +      NIR_PASS(progress, nir, nir_opt_dce);
> +      NIR_PASS(progress, nir, nir_opt_dead_cf);
> +   } while (progress);
> +
> +   NIR_PASS_V(nir, nir_remove_dead_variables, nir_var_local);
> +   NIR_PASS_V(nir, nir_convert_from_ssa, true);
> +
> +   /* Garbage collect dead instructions */
> +   nir_sweep(nir);
> +
> +   if (prog->dbgFlags & NV50_IR_DEBUG_BASIC)
> +      nir_print_shader(nir, stderr);
> +
>     return false;
>  }
>  
> -- 
> 2.14.3
> 
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20180311/5958eef0/attachment.sig>